Análise do desempenho operacional de uma caldeira de biomassa de leito fluidizado borbulhante

Abstract

The pulp and paper sector has a high energy demand, making the search for strategies that improve the operational performance of biomass boilers relevant for energy efficiency and sustainability. In this context, this study aimed to analyze the performance of a bubbling fluidized bed biomass boiler in a pulp and paper facility through the development of predictive models based on the eXtreme Gradient Boosting (XGBoost) algorithm, following the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology. Operational data were used as model inputs, including fuel-related variables, such as biomass consumption and moisture content, and boiler-related parameters, such as air supply rates and thermal conditions. Two regression models were developed: one to predict superheated steam flow rate and another to estimate carbon monoxide (CO) concentration in flue gases, used as an indicator of incomplete combustion. Hyperparameter optimization was performed using GridSearchCV, while model interpretation was conducted through the permutation importance technique, allowing the identification of the variables with the greatest influence on model behavior. The steam flow model demonstrated satisfactory predictive performance, with a Mean Absolute Percentage Error (MAPE) of 2.76% and a coefficient of determination (R2) of 0.87, indicating a good ability to represent steam generation under the evaluated conditions. The CO model showed more limited performance, with a Mean Absolute Error (MAE) of 2.23 ppm and an R2 of 0.61, reflecting the greater variability of this response variable. The integrated analysis showed that secondary air flow rate and post-booster air temperature were the most relevant variables for the steam production, whereas furnace and bed temperatures had the greatest influence on CO prediction. These two thermal variables were also among the main factors in the steam model, highlighting their relevance to both models. The findings highlight the potential of machine learning techniques in the pulp and paper industry by transforming operational data into useful information for process monitoring, pattern identification and decision-making support, with direct implication for efficiency and sustainability. Future work should focus on expanding the dataset in order to evaluate model robustness, generalization capacity and feasibility for application in operation monitoring.